2 * Copyright © 2014 Connor Abbott
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Connor Abbott (cwabbott0@gmail.com)
31 #include "util/hash_table.h"
32 #include "compiler/glsl/list.h"
33 #include "GL/gl.h" /* GLenum */
34 #include "util/list.h"
35 #include "util/ralloc.h"
37 #include "util/bitscan.h"
38 #include "util/bitset.h"
39 #include "util/macros.h"
40 #include "compiler/nir_types.h"
41 #include "compiler/shader_enums.h"
42 #include "compiler/shader_info.h"
46 #include "util/debug.h"
49 #include "nir_opcodes.h"
51 #if defined(_WIN32) && !defined(snprintf)
52 #define snprintf _snprintf
60 #define NIR_TRUE (~0u)
61 #define NIR_MAX_VEC_COMPONENTS 4
62 #define NIR_MAX_MATRIX_COLUMNS 4
63 typedef uint8_t nir_component_mask_t
;
65 /** Defines a cast function
67 * This macro defines a cast function from in_type to out_type where
68 * out_type is some structure type that contains a field of type out_type.
70 * Note that you have to be a bit careful as the generated cast function
73 #define NIR_DEFINE_CAST(name, in_type, out_type, field, \
74 type_field, type_value) \
75 static inline out_type * \
76 name(const in_type *parent) \
78 assert(parent && parent->type_field == type_value); \
79 return exec_node_data(out_type, parent, field); \
89 * Description of built-in state associated with a uniform
91 * \sa nir_variable::state_slots
94 gl_state_index16 tokens
[STATE_LENGTH
];
99 nir_var_shader_in
= (1 << 0),
100 nir_var_shader_out
= (1 << 1),
101 nir_var_shader_temp
= (1 << 2),
102 nir_var_function_temp
= (1 << 3),
103 nir_var_uniform
= (1 << 4),
104 nir_var_mem_ubo
= (1 << 5),
105 nir_var_system_value
= (1 << 6),
106 nir_var_mem_ssbo
= (1 << 7),
107 nir_var_mem_shared
= (1 << 8),
108 nir_var_mem_global
= (1 << 9),
116 nir_rounding_mode_undef
= 0,
117 nir_rounding_mode_rtne
= 1, /* round to nearest even */
118 nir_rounding_mode_ru
= 2, /* round up */
119 nir_rounding_mode_rd
= 3, /* round down */
120 nir_rounding_mode_rtz
= 4, /* round towards zero */
137 #define nir_const_value_to_array(arr, c, components, m) \
139 for (unsigned i = 0; i < components; ++i) \
143 typedef struct nir_constant
{
145 * Value of the constant.
147 * The field used to back the values supplied by the constant is determined
148 * by the type associated with the \c nir_variable. Constants may be
149 * scalars, vectors, or matrices.
151 nir_const_value values
[NIR_MAX_MATRIX_COLUMNS
][NIR_MAX_VEC_COMPONENTS
];
153 /* we could get this from the var->type but makes clone *much* easier to
154 * not have to care about the type.
156 unsigned num_elements
;
158 /* Array elements / Structure Fields */
159 struct nir_constant
**elements
;
163 * \brief Layout qualifiers for gl_FragDepth.
165 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
166 * with a layout qualifier.
169 nir_depth_layout_none
, /**< No depth layout is specified. */
170 nir_depth_layout_any
,
171 nir_depth_layout_greater
,
172 nir_depth_layout_less
,
173 nir_depth_layout_unchanged
177 * Enum keeping track of how a variable was declared.
181 * Normal declaration.
183 nir_var_declared_normally
= 0,
186 * Variable is implicitly generated by the compiler and should not be
187 * visible via the API.
190 } nir_var_declaration_type
;
193 * Either a uniform, global variable, shader input, or shader output. Based on
194 * ir_variable - it should be easy to translate between the two.
197 typedef struct nir_variable
{
198 struct exec_node node
;
201 * Declared type of the variable
203 const struct glsl_type
*type
;
206 * Declared name of the variable
210 struct nir_variable_data
{
212 * Storage class of the variable.
214 * \sa nir_variable_mode
216 nir_variable_mode mode
;
219 * Is the variable read-only?
221 * This is set for variables declared as \c const, shader inputs,
224 unsigned read_only
:1;
228 unsigned invariant
:1;
231 * When separate shader programs are enabled, only input/outputs between
232 * the stages of a multi-stage separate program can be safely removed
233 * from the shader interface. Other input/outputs must remains active.
235 * This is also used to make sure xfb varyings that are unused by the
236 * fragment shader are not removed.
238 unsigned always_active_io
:1;
241 * Interpolation mode for shader inputs / outputs
243 * \sa glsl_interp_mode
245 unsigned interpolation
:2;
248 * If non-zero, then this variable may be packed along with other variables
249 * into a single varying slot, so this offset should be applied when
250 * accessing components. For example, an offset of 1 means that the x
251 * component of this variable is actually stored in component y of the
252 * location specified by \c location.
254 unsigned location_frac
:2;
257 * If true, this variable represents an array of scalars that should
258 * be tightly packed. In other words, consecutive array elements
259 * should be stored one component apart, rather than one slot apart.
264 * Whether this is a fragment shader output implicitly initialized with
265 * the previous contents of the specified render target at the
266 * framebuffer location corresponding to this shader invocation.
268 unsigned fb_fetch_output
:1;
271 * Non-zero if this variable is considered bindless as defined by
272 * ARB_bindless_texture.
277 * Was an explicit binding set in the shader?
279 unsigned explicit_binding
:1;
282 * Was a transfer feedback buffer set in the shader?
284 unsigned explicit_xfb_buffer
:1;
287 * Was a transfer feedback stride set in the shader?
289 unsigned explicit_xfb_stride
:1;
292 * Was an explicit offset set in the shader?
294 unsigned explicit_offset
:1;
297 * \brief Layout qualifier for gl_FragDepth.
299 * This is not equal to \c ir_depth_layout_none if and only if this
300 * variable is \c gl_FragDepth and a layout qualifier is specified.
302 nir_depth_layout depth_layout
;
305 * Storage location of the base of this variable
307 * The precise meaning of this field depends on the nature of the variable.
309 * - Vertex shader input: one of the values from \c gl_vert_attrib.
310 * - Vertex shader output: one of the values from \c gl_varying_slot.
311 * - Geometry shader input: one of the values from \c gl_varying_slot.
312 * - Geometry shader output: one of the values from \c gl_varying_slot.
313 * - Fragment shader input: one of the values from \c gl_varying_slot.
314 * - Fragment shader output: one of the values from \c gl_frag_result.
315 * - Uniforms: Per-stage uniform slot number for default uniform block.
316 * - Uniforms: Index within the uniform block definition for UBO members.
317 * - Non-UBO Uniforms: uniform slot number.
318 * - Other: This field is not currently used.
320 * If the variable is a uniform, shader input, or shader output, and the
321 * slot has not been assigned, the value will be -1.
326 * The actual location of the variable in the IR. Only valid for inputs
329 unsigned int driver_location
;
332 * Vertex stream output identifier.
334 * For packed outputs, bit 31 is set and bits [2*i+1,2*i] indicate the
335 * stream of the i-th component.
340 * output index for dual source blending.
345 * Descriptor set binding for sampler or UBO.
350 * Initial binding point for a sampler or UBO.
352 * For array types, this represents the binding point for the first element.
357 * Location an atomic counter or transform feedback is stored at.
362 * Transform feedback buffer.
367 * Transform feedback stride.
372 * How the variable was declared. See nir_var_declaration_type.
374 * This is used to detect variables generated by the compiler, so should
375 * not be visible via the API.
377 unsigned how_declared
:2;
380 * ARB_shader_image_load_store qualifiers.
383 enum gl_access_qualifier access
;
385 /** Image internal format if specified explicitly, otherwise GL_NONE. */
391 * Built-in state that backs this uniform
393 * Once set at variable creation, \c state_slots must remain invariant.
394 * This is because, ideally, this array would be shared by all clones of
395 * this variable in the IR tree. In other words, we'd really like for it
396 * to be a fly-weight.
398 * If the variable is not a uniform, \c num_state_slots will be zero and
399 * \c state_slots will be \c NULL.
402 unsigned num_state_slots
; /**< Number of state slots used */
403 nir_state_slot
*state_slots
; /**< State descriptors. */
407 * Constant expression assigned in the initializer of the variable
409 * This field should only be used temporarily by creators of NIR shaders
410 * and then lower_constant_initializers can be used to get rid of them.
411 * Most of the rest of NIR ignores this field or asserts that it's NULL.
413 nir_constant
*constant_initializer
;
416 * For variables that are in an interface block or are an instance of an
417 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
419 * \sa ir_variable::location
421 const struct glsl_type
*interface_type
;
424 * Description of per-member data for per-member struct variables
426 * This is used for variables which are actually an amalgamation of
427 * multiple entities such as a struct of built-in values or a struct of
428 * inputs each with their own layout specifier. This is only allowed on
429 * variables with a struct or array of array of struct type.
431 unsigned num_members
;
432 struct nir_variable_data
*members
;
435 #define nir_foreach_variable(var, var_list) \
436 foreach_list_typed(nir_variable, var, node, var_list)
438 #define nir_foreach_variable_safe(var, var_list) \
439 foreach_list_typed_safe(nir_variable, var, node, var_list)
442 nir_variable_is_global(const nir_variable
*var
)
444 return var
->data
.mode
!= nir_var_function_temp
;
447 typedef struct nir_register
{
448 struct exec_node node
;
450 unsigned num_components
; /** < number of vector components */
451 unsigned num_array_elems
; /** < size of array (0 for no array) */
453 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
456 /** generic register index. */
459 /** only for debug purposes, can be NULL */
462 /** set of nir_srcs where this register is used (read from) */
463 struct list_head uses
;
465 /** set of nir_dests where this register is defined (written to) */
466 struct list_head defs
;
468 /** set of nir_ifs where this register is used as a condition */
469 struct list_head if_uses
;
472 #define nir_foreach_register(reg, reg_list) \
473 foreach_list_typed(nir_register, reg, node, reg_list)
474 #define nir_foreach_register_safe(reg, reg_list) \
475 foreach_list_typed_safe(nir_register, reg, node, reg_list)
477 typedef enum PACKED
{
479 nir_instr_type_deref
,
482 nir_instr_type_intrinsic
,
483 nir_instr_type_load_const
,
485 nir_instr_type_ssa_undef
,
487 nir_instr_type_parallel_copy
,
490 typedef struct nir_instr
{
491 struct exec_node node
;
492 struct nir_block
*block
;
495 /* A temporary for optimization and analysis passes to use for storing
496 * flags. For instance, DCE uses this to store the "dead/live" info.
500 /** generic instruction index. */
504 static inline nir_instr
*
505 nir_instr_next(nir_instr
*instr
)
507 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
508 if (exec_node_is_tail_sentinel(next
))
511 return exec_node_data(nir_instr
, next
, node
);
514 static inline nir_instr
*
515 nir_instr_prev(nir_instr
*instr
)
517 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
518 if (exec_node_is_head_sentinel(prev
))
521 return exec_node_data(nir_instr
, prev
, node
);
525 nir_instr_is_first(const nir_instr
*instr
)
527 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr
->node
));
531 nir_instr_is_last(const nir_instr
*instr
)
533 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr
->node
));
536 typedef struct nir_ssa_def
{
537 /** for debugging only, can be NULL */
540 /** generic SSA definition index. */
543 /** Index into the live_in and live_out bitfields */
546 /** Instruction which produces this SSA value. */
547 nir_instr
*parent_instr
;
549 /** set of nir_instrs where this register is used (read from) */
550 struct list_head uses
;
552 /** set of nir_ifs where this register is used as a condition */
553 struct list_head if_uses
;
555 uint8_t num_components
;
557 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
565 struct nir_src
*indirect
; /** < NULL for no indirect offset */
566 unsigned base_offset
;
568 /* TODO use-def chain goes here */
572 nir_instr
*parent_instr
;
573 struct list_head def_link
;
576 struct nir_src
*indirect
; /** < NULL for no indirect offset */
577 unsigned base_offset
;
579 /* TODO def-use chain goes here */
584 typedef struct nir_src
{
586 /** Instruction that consumes this value as a source. */
587 nir_instr
*parent_instr
;
588 struct nir_if
*parent_if
;
591 struct list_head use_link
;
601 static inline nir_src
604 nir_src src
= { { NULL
} };
608 #define NIR_SRC_INIT nir_src_init()
610 #define nir_foreach_use(src, reg_or_ssa_def) \
611 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
613 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
614 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
616 #define nir_foreach_if_use(src, reg_or_ssa_def) \
617 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
619 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
620 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
631 static inline nir_dest
634 nir_dest dest
= { { { NULL
} } };
638 #define NIR_DEST_INIT nir_dest_init()
640 #define nir_foreach_def(dest, reg) \
641 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
643 #define nir_foreach_def_safe(dest, reg) \
644 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
646 static inline nir_src
647 nir_src_for_ssa(nir_ssa_def
*def
)
649 nir_src src
= NIR_SRC_INIT
;
657 static inline nir_src
658 nir_src_for_reg(nir_register
*reg
)
660 nir_src src
= NIR_SRC_INIT
;
664 src
.reg
.indirect
= NULL
;
665 src
.reg
.base_offset
= 0;
670 static inline nir_dest
671 nir_dest_for_reg(nir_register
*reg
)
673 nir_dest dest
= NIR_DEST_INIT
;
680 static inline unsigned
681 nir_src_bit_size(nir_src src
)
683 return src
.is_ssa
? src
.ssa
->bit_size
: src
.reg
.reg
->bit_size
;
686 static inline unsigned
687 nir_src_num_components(nir_src src
)
689 return src
.is_ssa
? src
.ssa
->num_components
: src
.reg
.reg
->num_components
;
693 nir_src_is_const(nir_src src
)
696 src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
;
699 int64_t nir_src_as_int(nir_src src
);
700 uint64_t nir_src_as_uint(nir_src src
);
701 bool nir_src_as_bool(nir_src src
);
702 double nir_src_as_float(nir_src src
);
703 int64_t nir_src_comp_as_int(nir_src src
, unsigned component
);
704 uint64_t nir_src_comp_as_uint(nir_src src
, unsigned component
);
705 bool nir_src_comp_as_bool(nir_src src
, unsigned component
);
706 double nir_src_comp_as_float(nir_src src
, unsigned component
);
708 static inline unsigned
709 nir_dest_bit_size(nir_dest dest
)
711 return dest
.is_ssa
? dest
.ssa
.bit_size
: dest
.reg
.reg
->bit_size
;
714 static inline unsigned
715 nir_dest_num_components(nir_dest dest
)
717 return dest
.is_ssa
? dest
.ssa
.num_components
: dest
.reg
.reg
->num_components
;
720 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
721 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
727 * \name input modifiers
731 * For inputs interpreted as floating point, flips the sign bit. For
732 * inputs interpreted as integers, performs the two's complement negation.
737 * Clears the sign bit for floating point values, and computes the integer
738 * absolute value for integers. Note that the negate modifier acts after
739 * the absolute value modifier, therefore if both are set then all inputs
740 * will become negative.
746 * For each input component, says which component of the register it is
747 * chosen from. Note that which elements of the swizzle are used and which
748 * are ignored are based on the write mask for most opcodes - for example,
749 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
750 * a swizzle of {2, x, 1, 0} where x means "don't care."
752 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
759 * \name saturate output modifier
761 * Only valid for opcodes that output floating-point numbers. Clamps the
762 * output to between 0.0 and 1.0 inclusive.
767 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
770 /** NIR sized and unsized types
772 * The values in this enum are carefully chosen so that the sized type is
773 * just the unsized type OR the number of bits.
776 nir_type_invalid
= 0, /* Not a valid type */
780 nir_type_float
= 128,
781 nir_type_bool1
= 1 | nir_type_bool
,
782 nir_type_bool32
= 32 | nir_type_bool
,
783 nir_type_int1
= 1 | nir_type_int
,
784 nir_type_int8
= 8 | nir_type_int
,
785 nir_type_int16
= 16 | nir_type_int
,
786 nir_type_int32
= 32 | nir_type_int
,
787 nir_type_int64
= 64 | nir_type_int
,
788 nir_type_uint1
= 1 | nir_type_uint
,
789 nir_type_uint8
= 8 | nir_type_uint
,
790 nir_type_uint16
= 16 | nir_type_uint
,
791 nir_type_uint32
= 32 | nir_type_uint
,
792 nir_type_uint64
= 64 | nir_type_uint
,
793 nir_type_float16
= 16 | nir_type_float
,
794 nir_type_float32
= 32 | nir_type_float
,
795 nir_type_float64
= 64 | nir_type_float
,
798 #define NIR_ALU_TYPE_SIZE_MASK 0x79
799 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
801 static inline unsigned
802 nir_alu_type_get_type_size(nir_alu_type type
)
804 return type
& NIR_ALU_TYPE_SIZE_MASK
;
807 static inline unsigned
808 nir_alu_type_get_base_type(nir_alu_type type
)
810 return type
& NIR_ALU_TYPE_BASE_TYPE_MASK
;
813 static inline nir_alu_type
814 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
818 return nir_type_bool1
;
821 return nir_type_uint32
;
824 return nir_type_int32
;
826 case GLSL_TYPE_UINT16
:
827 return nir_type_uint16
;
829 case GLSL_TYPE_INT16
:
830 return nir_type_int16
;
832 case GLSL_TYPE_UINT8
:
833 return nir_type_uint8
;
835 return nir_type_int8
;
836 case GLSL_TYPE_UINT64
:
837 return nir_type_uint64
;
839 case GLSL_TYPE_INT64
:
840 return nir_type_int64
;
842 case GLSL_TYPE_FLOAT
:
843 return nir_type_float32
;
845 case GLSL_TYPE_FLOAT16
:
846 return nir_type_float16
;
848 case GLSL_TYPE_DOUBLE
:
849 return nir_type_float64
;
852 unreachable("unknown type");
856 static inline nir_alu_type
857 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
859 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
862 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
863 nir_rounding_mode rnd
);
866 nir_op_vec(unsigned components
)
868 switch (components
) {
869 case 1: return nir_op_mov
;
870 case 2: return nir_op_vec2
;
871 case 3: return nir_op_vec3
;
872 case 4: return nir_op_vec4
;
873 default: unreachable("bad component count");
879 * Operation where the first two sources are commutative.
881 * For 2-source operations, this just mathematical commutativity. Some
882 * 3-source operations, like ffma, are only commutative in the first two
885 NIR_OP_IS_2SRC_COMMUTATIVE
= (1 << 0),
886 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
887 } nir_op_algebraic_property
;
895 * The number of components in the output
897 * If non-zero, this is the size of the output and input sizes are
898 * explicitly given; swizzle and writemask are still in effect, but if
899 * the output component is masked out, then the input component may
902 * If zero, the opcode acts in the standard, per-component manner; the
903 * operation is performed on each component (except the ones that are
904 * masked out) with the input being taken from the input swizzle for
907 * The size of some of the inputs may be given (i.e. non-zero) even
908 * though output_size is zero; in that case, the inputs with a zero
909 * size act per-component, while the inputs with non-zero size don't.
911 unsigned output_size
;
914 * The type of vector that the instruction outputs. Note that the
915 * staurate modifier is only allowed on outputs with the float type.
918 nir_alu_type output_type
;
921 * The number of components in each input
923 unsigned input_sizes
[NIR_MAX_VEC_COMPONENTS
];
926 * The type of vector that each input takes. Note that negate and
927 * absolute value are only allowed on inputs with int or float type and
928 * behave differently on the two.
930 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
932 nir_op_algebraic_property algebraic_properties
;
934 /* Whether this represents a numeric conversion opcode */
938 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
940 typedef struct nir_alu_instr
{
944 /** Indicates that this ALU instruction generates an exact value
946 * This is kind of a mixture of GLSL "precise" and "invariant" and not
947 * really equivalent to either. This indicates that the value generated by
948 * this operation is high-precision and any code transformations that touch
949 * it must ensure that the resulting value is bit-for-bit identical to the
958 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
959 nir_alu_instr
*instr
);
960 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
961 nir_alu_instr
*instr
);
963 /* is this source channel used? */
965 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
968 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
969 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
971 return (instr
->dest
.write_mask
>> channel
) & 1;
974 static inline nir_component_mask_t
975 nir_alu_instr_src_read_mask(const nir_alu_instr
*instr
, unsigned src
)
977 nir_component_mask_t read_mask
= 0;
978 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; c
++) {
979 if (!nir_alu_instr_channel_used(instr
, src
, c
))
982 read_mask
|= (1 << instr
->src
[src
].swizzle
[c
]);
988 * For instructions whose destinations are SSA, get the number of channels
991 static inline unsigned
992 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
994 assert(instr
->dest
.dest
.is_ssa
);
996 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
997 return nir_op_infos
[instr
->op
].input_sizes
[src
];
999 return instr
->dest
.dest
.ssa
.num_components
;
1002 bool nir_const_value_negative_equal(const nir_const_value
*c1
,
1003 const nir_const_value
*c2
,
1004 unsigned components
,
1005 nir_alu_type base_type
,
1008 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
1009 unsigned src1
, unsigned src2
);
1011 bool nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
1012 const nir_alu_instr
*alu2
,
1013 unsigned src1
, unsigned src2
);
1017 nir_deref_type_array
,
1018 nir_deref_type_array_wildcard
,
1019 nir_deref_type_ptr_as_array
,
1020 nir_deref_type_struct
,
1021 nir_deref_type_cast
,
1027 /** The type of this deref instruction */
1028 nir_deref_type deref_type
;
1030 /** The mode of the underlying variable */
1031 nir_variable_mode mode
;
1033 /** The dereferenced type of the resulting pointer value */
1034 const struct glsl_type
*type
;
1037 /** Variable being dereferenced if deref_type is a deref_var */
1040 /** Parent deref if deref_type is not deref_var */
1044 /** Additional deref parameters */
1055 unsigned ptr_stride
;
1059 /** Destination to store the resulting "pointer" */
1063 static inline nir_deref_instr
*nir_src_as_deref(nir_src src
);
1065 static inline nir_deref_instr
*
1066 nir_deref_instr_parent(const nir_deref_instr
*instr
)
1068 if (instr
->deref_type
== nir_deref_type_var
)
1071 return nir_src_as_deref(instr
->parent
);
1074 static inline nir_variable
*
1075 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1077 while (instr
->deref_type
!= nir_deref_type_var
) {
1078 if (instr
->deref_type
== nir_deref_type_cast
)
1081 instr
= nir_deref_instr_parent(instr
);
1087 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1088 bool nir_deref_instr_has_complex_use(nir_deref_instr
*instr
);
1090 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1092 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*instr
);
1097 struct nir_function
*callee
;
1099 unsigned num_params
;
1103 #include "nir_intrinsics.h"
1105 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1107 /** Represents an intrinsic
1109 * An intrinsic is an instruction type for handling things that are
1110 * more-or-less regular operations but don't just consume and produce SSA
1111 * values like ALU operations do. Intrinsics are not for things that have
1112 * special semantic meaning such as phi nodes and parallel copies.
1113 * Examples of intrinsics include variable load/store operations, system
1114 * value loads, and the like. Even though texturing more-or-less falls
1115 * under this category, texturing is its own instruction type because
1116 * trying to represent texturing with intrinsics would lead to a
1117 * combinatorial explosion of intrinsic opcodes.
1119 * By having a single instruction type for handling a lot of different
1120 * cases, optimization passes can look for intrinsics and, for the most
1121 * part, completely ignore them. Each intrinsic type also has a few
1122 * possible flags that govern whether or not they can be reordered or
1123 * eliminated. That way passes like dead code elimination can still work
1124 * on intrisics without understanding the meaning of each.
1126 * Each intrinsic has some number of constant indices, some number of
1127 * variables, and some number of sources. What these sources, variables,
1128 * and indices mean depends on the intrinsic and is documented with the
1129 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1130 * instructions are the only types of instruction that can operate on
1136 nir_intrinsic_op intrinsic
;
1140 /** number of components if this is a vectorized intrinsic
1142 * Similarly to ALU operations, some intrinsics are vectorized.
1143 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1144 * For vectorized intrinsics, the num_components field specifies the
1145 * number of destination components and the number of source components
1146 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1148 uint8_t num_components
;
1150 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1153 } nir_intrinsic_instr
;
1155 static inline nir_variable
*
1156 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1158 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1162 * \name NIR intrinsics semantic flags
1164 * information about what the compiler can do with the intrinsics.
1166 * \sa nir_intrinsic_info::flags
1170 * whether the intrinsic can be safely eliminated if none of its output
1171 * value is not being used.
1173 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1176 * Whether the intrinsic can be reordered with respect to any other
1177 * intrinsic, i.e. whether the only reordering dependencies of the
1178 * intrinsic are due to the register reads/writes.
1180 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1181 } nir_intrinsic_semantic_flag
;
1184 * \name NIR intrinsics const-index flag
1186 * Indicates the usage of a const_index slot.
1188 * \sa nir_intrinsic_info::index_map
1192 * Generally instructions that take a offset src argument, can encode
1193 * a constant 'base' value which is added to the offset.
1195 NIR_INTRINSIC_BASE
= 1,
1198 * For store instructions, a writemask for the store.
1200 NIR_INTRINSIC_WRMASK
= 2,
1203 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1205 NIR_INTRINSIC_STREAM_ID
= 3,
1208 * The clip-plane id for load_user_clip_plane intrinsic.
1210 NIR_INTRINSIC_UCP_ID
= 4,
1213 * The amount of data, starting from BASE, that this instruction may
1214 * access. This is used to provide bounds if the offset is not constant.
1216 NIR_INTRINSIC_RANGE
= 5,
1219 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1221 NIR_INTRINSIC_DESC_SET
= 6,
1224 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1226 NIR_INTRINSIC_BINDING
= 7,
1231 NIR_INTRINSIC_COMPONENT
= 8,
1234 * Interpolation mode (only meaningful for FS inputs).
1236 NIR_INTRINSIC_INTERP_MODE
= 9,
1239 * A binary nir_op to use when performing a reduction or scan operation
1241 NIR_INTRINSIC_REDUCTION_OP
= 10,
1244 * Cluster size for reduction operations
1246 NIR_INTRINSIC_CLUSTER_SIZE
= 11,
1249 * Parameter index for a load_param intrinsic
1251 NIR_INTRINSIC_PARAM_IDX
= 12,
1254 * Image dimensionality for image intrinsics
1256 * One of GLSL_SAMPLER_DIM_*
1258 NIR_INTRINSIC_IMAGE_DIM
= 13,
1261 * Non-zero if we are accessing an array image
1263 NIR_INTRINSIC_IMAGE_ARRAY
= 14,
1266 * Image format for image intrinsics
1268 NIR_INTRINSIC_FORMAT
= 15,
1271 * Access qualifiers for image and memory access intrinsics
1273 NIR_INTRINSIC_ACCESS
= 16,
1276 * Alignment for offsets and addresses
1278 * These two parameters, specify an alignment in terms of a multiplier and
1279 * an offset. The offset or address parameter X of the intrinsic is
1280 * guaranteed to satisfy the following:
1282 * (X - align_offset) % align_mul == 0
1284 NIR_INTRINSIC_ALIGN_MUL
= 17,
1285 NIR_INTRINSIC_ALIGN_OFFSET
= 18,
1288 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1290 NIR_INTRINSIC_DESC_TYPE
= 19,
1292 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1294 } nir_intrinsic_index_flag
;
1296 #define NIR_INTRINSIC_MAX_INPUTS 5
1301 unsigned num_srcs
; /** < number of register/SSA inputs */
1303 /** number of components of each input register
1305 * If this value is 0, the number of components is given by the
1306 * num_components field of nir_intrinsic_instr. If this value is -1, the
1307 * intrinsic consumes however many components are provided and it is not
1310 int src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1314 /** number of components of the output register
1316 * If this value is 0, the number of components is given by the
1317 * num_components field of nir_intrinsic_instr.
1319 unsigned dest_components
;
1321 /** bitfield of legal bit sizes */
1322 unsigned dest_bit_sizes
;
1324 /** the number of constant indices used by the intrinsic */
1325 unsigned num_indices
;
1327 /** indicates the usage of intr->const_index[n] */
1328 unsigned index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1330 /** semantic flags for calls to this intrinsic */
1331 nir_intrinsic_semantic_flag flags
;
1332 } nir_intrinsic_info
;
1334 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1336 static inline unsigned
1337 nir_intrinsic_src_components(nir_intrinsic_instr
*intr
, unsigned srcn
)
1339 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1340 assert(srcn
< info
->num_srcs
);
1341 if (info
->src_components
[srcn
] > 0)
1342 return info
->src_components
[srcn
];
1343 else if (info
->src_components
[srcn
] == 0)
1344 return intr
->num_components
;
1346 return nir_src_num_components(intr
->src
[srcn
]);
1349 static inline unsigned
1350 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1352 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1353 if (!info
->has_dest
)
1355 else if (info
->dest_components
)
1356 return info
->dest_components
;
1358 return intr
->num_components
;
1361 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1362 static inline type \
1363 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1365 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1366 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1367 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1369 static inline void \
1370 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1372 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1373 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1374 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1377 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1378 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1379 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1380 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1381 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1382 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1383 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1384 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1385 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1386 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1387 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1388 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1389 INTRINSIC_IDX_ACCESSORS(image_dim
, IMAGE_DIM
, enum glsl_sampler_dim
)
1390 INTRINSIC_IDX_ACCESSORS(image_array
, IMAGE_ARRAY
, bool)
1391 INTRINSIC_IDX_ACCESSORS(access
, ACCESS
, enum gl_access_qualifier
)
1392 INTRINSIC_IDX_ACCESSORS(format
, FORMAT
, unsigned)
1393 INTRINSIC_IDX_ACCESSORS(align_mul
, ALIGN_MUL
, unsigned)
1394 INTRINSIC_IDX_ACCESSORS(align_offset
, ALIGN_OFFSET
, unsigned)
1395 INTRINSIC_IDX_ACCESSORS(desc_type
, DESC_TYPE
, unsigned)
1398 nir_intrinsic_set_align(nir_intrinsic_instr
*intrin
,
1399 unsigned align_mul
, unsigned align_offset
)
1401 assert(util_is_power_of_two_nonzero(align_mul
));
1402 assert(align_offset
< align_mul
);
1403 nir_intrinsic_set_align_mul(intrin
, align_mul
);
1404 nir_intrinsic_set_align_offset(intrin
, align_offset
);
1407 /** Returns a simple alignment for a load/store intrinsic offset
1409 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1410 * and ALIGN_OFFSET parameters, this helper takes both into account and
1411 * provides a single simple alignment parameter. The offset X is guaranteed
1412 * to satisfy X % align == 0.
1414 static inline unsigned
1415 nir_intrinsic_align(const nir_intrinsic_instr
*intrin
)
1417 const unsigned align_mul
= nir_intrinsic_align_mul(intrin
);
1418 const unsigned align_offset
= nir_intrinsic_align_offset(intrin
);
1419 assert(align_offset
< align_mul
);
1420 return align_offset
? 1 << (ffs(align_offset
) - 1) : align_mul
;
1423 /* Converts a image_deref_* intrinsic into a image_* one */
1424 void nir_rewrite_image_intrinsic(nir_intrinsic_instr
*instr
,
1425 nir_ssa_def
*handle
, bool bindless
);
1428 * \group texture information
1430 * This gives semantic information about textures which is useful to the
1431 * frontend, the backend, and lowering passes, but not the optimizer.
1436 nir_tex_src_projector
,
1437 nir_tex_src_comparator
, /* shadow comparator */
1441 nir_tex_src_min_lod
,
1442 nir_tex_src_ms_index
, /* MSAA sample index */
1443 nir_tex_src_ms_mcs
, /* MSAA compression value */
1446 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
1447 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
1448 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
1449 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
1450 nir_tex_src_texture_handle
, /* < bindless texture handle */
1451 nir_tex_src_sampler_handle
, /* < bindless sampler handle */
1452 nir_tex_src_plane
, /* < selects plane for planar textures */
1453 nir_num_tex_src_types
1458 nir_tex_src_type src_type
;
1462 nir_texop_tex
, /**< Regular texture look-up */
1463 nir_texop_txb
, /**< Texture look-up with LOD bias */
1464 nir_texop_txl
, /**< Texture look-up with explicit LOD */
1465 nir_texop_txd
, /**< Texture look-up with partial derivatives */
1466 nir_texop_txf
, /**< Texel fetch with explicit LOD */
1467 nir_texop_txf_ms
, /**< Multisample texture fetch */
1468 nir_texop_txf_ms_fb
, /**< Multisample texture fetch from framebuffer */
1469 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
1470 nir_texop_txs
, /**< Texture size */
1471 nir_texop_lod
, /**< Texture lod query */
1472 nir_texop_tg4
, /**< Texture gather */
1473 nir_texop_query_levels
, /**< Texture levels query */
1474 nir_texop_texture_samples
, /**< Texture samples query */
1475 nir_texop_samples_identical
, /**< Query whether all samples are definitely
1483 enum glsl_sampler_dim sampler_dim
;
1484 nir_alu_type dest_type
;
1489 unsigned num_srcs
, coord_components
;
1490 bool is_array
, is_shadow
;
1493 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1494 * components or the new-style shadow that outputs 1 component.
1496 bool is_new_style_shadow
;
1498 /* gather component selector */
1499 unsigned component
: 2;
1501 /* gather offsets */
1502 int8_t tg4_offsets
[4][2];
1504 /* True if the texture index or handle is not dynamically uniform */
1505 bool texture_non_uniform
;
1507 /* True if the sampler index or handle is not dynamically uniform */
1508 bool sampler_non_uniform
;
1510 /** The texture index
1512 * If this texture instruction has a nir_tex_src_texture_offset source,
1513 * then the texture index is given by texture_index + texture_offset.
1515 unsigned texture_index
;
1517 /** The size of the texture array or 0 if it's not an array */
1518 unsigned texture_array_size
;
1520 /** The sampler index
1522 * The following operations do not require a sampler and, as such, this
1523 * field should be ignored:
1525 * - nir_texop_txf_ms
1528 * - nir_texop_query_levels
1529 * - nir_texop_texture_samples
1530 * - nir_texop_samples_identical
1532 * If this texture instruction has a nir_tex_src_sampler_offset source,
1533 * then the sampler index is given by sampler_index + sampler_offset.
1535 unsigned sampler_index
;
1538 static inline unsigned
1539 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
1541 switch (instr
->op
) {
1542 case nir_texop_txs
: {
1544 switch (instr
->sampler_dim
) {
1545 case GLSL_SAMPLER_DIM_1D
:
1546 case GLSL_SAMPLER_DIM_BUF
:
1549 case GLSL_SAMPLER_DIM_2D
:
1550 case GLSL_SAMPLER_DIM_CUBE
:
1551 case GLSL_SAMPLER_DIM_MS
:
1552 case GLSL_SAMPLER_DIM_RECT
:
1553 case GLSL_SAMPLER_DIM_EXTERNAL
:
1554 case GLSL_SAMPLER_DIM_SUBPASS
:
1557 case GLSL_SAMPLER_DIM_3D
:
1561 unreachable("not reached");
1563 if (instr
->is_array
)
1571 case nir_texop_texture_samples
:
1572 case nir_texop_query_levels
:
1573 case nir_texop_samples_identical
:
1577 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
1584 /* Returns true if this texture operation queries something about the texture
1585 * rather than actually sampling it.
1588 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
1590 switch (instr
->op
) {
1593 case nir_texop_texture_samples
:
1594 case nir_texop_query_levels
:
1595 case nir_texop_txf_ms_mcs
:
1602 case nir_texop_txf_ms
:
1603 case nir_texop_txf_ms_fb
:
1607 unreachable("Invalid texture opcode");
1612 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1614 switch (instr
->op
) {
1635 static inline nir_alu_type
1636 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
1638 switch (instr
->src
[src
].src_type
) {
1639 case nir_tex_src_coord
:
1640 switch (instr
->op
) {
1642 case nir_texop_txf_ms
:
1643 case nir_texop_txf_ms_fb
:
1644 case nir_texop_txf_ms_mcs
:
1645 case nir_texop_samples_identical
:
1646 return nir_type_int
;
1649 return nir_type_float
;
1652 case nir_tex_src_lod
:
1653 switch (instr
->op
) {
1656 return nir_type_int
;
1659 return nir_type_float
;
1662 case nir_tex_src_projector
:
1663 case nir_tex_src_comparator
:
1664 case nir_tex_src_bias
:
1665 case nir_tex_src_ddx
:
1666 case nir_tex_src_ddy
:
1667 return nir_type_float
;
1669 case nir_tex_src_offset
:
1670 case nir_tex_src_ms_index
:
1671 case nir_tex_src_texture_offset
:
1672 case nir_tex_src_sampler_offset
:
1673 return nir_type_int
;
1676 unreachable("Invalid texture source type");
1680 static inline unsigned
1681 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
1683 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
1684 return instr
->coord_components
;
1686 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
1687 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
1690 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
1691 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
1692 if (instr
->is_array
)
1693 return instr
->coord_components
- 1;
1695 return instr
->coord_components
;
1698 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
1699 * the offset, since a cube maps to a single face.
1701 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
1702 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
1704 else if (instr
->is_array
)
1705 return instr
->coord_components
- 1;
1707 return instr
->coord_components
;
1714 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
1716 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
1717 if (instr
->src
[i
].src_type
== type
)
1723 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
1724 nir_tex_src_type src_type
,
1727 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
1729 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
);
1736 nir_const_value value
[];
1737 } nir_load_const_instr
;
1739 #define nir_const_load_to_arr(arr, l, m) \
1741 nir_const_value_to_array(arr, l->value, l->def.num_components, m); \
1755 /* creates a new SSA variable in an undefined state */
1760 } nir_ssa_undef_instr
;
1763 struct exec_node node
;
1765 /* The predecessor block corresponding to this source */
1766 struct nir_block
*pred
;
1771 #define nir_foreach_phi_src(phi_src, phi) \
1772 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
1773 #define nir_foreach_phi_src_safe(phi_src, phi) \
1774 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
1779 struct exec_list srcs
; /** < list of nir_phi_src */
1785 struct exec_node node
;
1788 } nir_parallel_copy_entry
;
1790 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
1791 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
1796 /* A list of nir_parallel_copy_entrys. The sources of all of the
1797 * entries are copied to the corresponding destinations "in parallel".
1798 * In other words, if we have two entries: a -> b and b -> a, the values
1801 struct exec_list entries
;
1802 } nir_parallel_copy_instr
;
1804 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
1805 type
, nir_instr_type_alu
)
1806 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
1807 type
, nir_instr_type_deref
)
1808 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
1809 type
, nir_instr_type_call
)
1810 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
1811 type
, nir_instr_type_jump
)
1812 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
1813 type
, nir_instr_type_tex
)
1814 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
1815 type
, nir_instr_type_intrinsic
)
1816 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
1817 type
, nir_instr_type_load_const
)
1818 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
1819 type
, nir_instr_type_ssa_undef
)
1820 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
1821 type
, nir_instr_type_phi
)
1822 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
1823 nir_parallel_copy_instr
, instr
,
1824 type
, nir_instr_type_parallel_copy
)
1829 * Control flow consists of a tree of control flow nodes, which include
1830 * if-statements and loops. The leaves of the tree are basic blocks, lists of
1831 * instructions that always run start-to-finish. Each basic block also keeps
1832 * track of its successors (blocks which may run immediately after the current
1833 * block) and predecessors (blocks which could have run immediately before the
1834 * current block). Each function also has a start block and an end block which
1835 * all return statements point to (which is always empty). Together, all the
1836 * blocks with their predecessors and successors make up the control flow
1837 * graph (CFG) of the function. There are helpers that modify the tree of
1838 * control flow nodes while modifying the CFG appropriately; these should be
1839 * used instead of modifying the tree directly.
1846 nir_cf_node_function
1849 typedef struct nir_cf_node
{
1850 struct exec_node node
;
1851 nir_cf_node_type type
;
1852 struct nir_cf_node
*parent
;
1855 typedef struct nir_block
{
1856 nir_cf_node cf_node
;
1858 struct exec_list instr_list
; /** < list of nir_instr */
1860 /** generic block index; generated by nir_index_blocks */
1864 * Each block can only have up to 2 successors, so we put them in a simple
1865 * array - no need for anything more complicated.
1867 struct nir_block
*successors
[2];
1869 /* Set of nir_block predecessors in the CFG */
1870 struct set
*predecessors
;
1873 * this node's immediate dominator in the dominance tree - set to NULL for
1876 struct nir_block
*imm_dom
;
1878 /* This node's children in the dominance tree */
1879 unsigned num_dom_children
;
1880 struct nir_block
**dom_children
;
1882 /* Set of nir_blocks on the dominance frontier of this block */
1883 struct set
*dom_frontier
;
1886 * These two indices have the property that dom_{pre,post}_index for each
1887 * child of this block in the dominance tree will always be between
1888 * dom_pre_index and dom_post_index for this block, which makes testing if
1889 * a given block is dominated by another block an O(1) operation.
1891 unsigned dom_pre_index
, dom_post_index
;
1893 /* live in and out for this block; used for liveness analysis */
1894 BITSET_WORD
*live_in
;
1895 BITSET_WORD
*live_out
;
1898 static inline nir_instr
*
1899 nir_block_first_instr(nir_block
*block
)
1901 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
1902 return exec_node_data(nir_instr
, head
, node
);
1905 static inline nir_instr
*
1906 nir_block_last_instr(nir_block
*block
)
1908 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
1909 return exec_node_data(nir_instr
, tail
, node
);
1913 nir_block_ends_in_jump(nir_block
*block
)
1915 return !exec_list_is_empty(&block
->instr_list
) &&
1916 nir_block_last_instr(block
)->type
== nir_instr_type_jump
;
1919 #define nir_foreach_instr(instr, block) \
1920 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
1921 #define nir_foreach_instr_reverse(instr, block) \
1922 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
1923 #define nir_foreach_instr_safe(instr, block) \
1924 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
1925 #define nir_foreach_instr_reverse_safe(instr, block) \
1926 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
1929 nir_selection_control_none
= 0x0,
1930 nir_selection_control_flatten
= 0x1,
1931 nir_selection_control_dont_flatten
= 0x2,
1932 } nir_selection_control
;
1934 typedef struct nir_if
{
1935 nir_cf_node cf_node
;
1937 nir_selection_control control
;
1939 struct exec_list then_list
; /** < list of nir_cf_node */
1940 struct exec_list else_list
; /** < list of nir_cf_node */
1946 /** Instruction that generates nif::condition. */
1947 nir_instr
*conditional_instr
;
1949 /** Block within ::nif that has the break instruction. */
1950 nir_block
*break_block
;
1952 /** Last block for the then- or else-path that does not contain the break. */
1953 nir_block
*continue_from_block
;
1955 /** True when ::break_block is in the else-path of ::nif. */
1956 bool continue_from_then
;
1959 /* This is true if the terminators exact trip count is unknown. For
1962 * for (int i = 0; i < imin(x, 4); i++)
1965 * Here loop analysis would have set a max_trip_count of 4 however we dont
1966 * know for sure that this is the exact trip count.
1968 bool exact_trip_count_unknown
;
1970 struct list_head loop_terminator_link
;
1971 } nir_loop_terminator
;
1974 /* Estimated cost (in number of instructions) of the loop */
1975 unsigned instr_cost
;
1977 /* Guessed trip count based on array indexing */
1978 unsigned guessed_trip_count
;
1980 /* Maximum number of times the loop is run (if known) */
1981 unsigned max_trip_count
;
1983 /* Do we know the exact number of times the loop will be run */
1984 bool exact_trip_count_known
;
1986 /* Unroll the loop regardless of its size */
1989 /* Does the loop contain complex loop terminators, continues or other
1990 * complex behaviours? If this is true we can't rely on
1991 * loop_terminator_list to be complete or accurate.
1995 nir_loop_terminator
*limiting_terminator
;
1997 /* A list of loop_terminators terminating this loop. */
1998 struct list_head loop_terminator_list
;
2002 nir_loop_control_none
= 0x0,
2003 nir_loop_control_unroll
= 0x1,
2004 nir_loop_control_dont_unroll
= 0x2,
2008 nir_cf_node cf_node
;
2010 struct exec_list body
; /** < list of nir_cf_node */
2012 nir_loop_info
*info
;
2013 nir_loop_control control
;
2014 bool partially_unrolled
;
2018 * Various bits of metadata that can may be created or required by
2019 * optimization and analysis passes
2022 nir_metadata_none
= 0x0,
2023 nir_metadata_block_index
= 0x1,
2024 nir_metadata_dominance
= 0x2,
2025 nir_metadata_live_ssa_defs
= 0x4,
2026 nir_metadata_not_properly_reset
= 0x8,
2027 nir_metadata_loop_analysis
= 0x10,
2031 nir_cf_node cf_node
;
2033 /** pointer to the function of which this is an implementation */
2034 struct nir_function
*function
;
2036 struct exec_list body
; /** < list of nir_cf_node */
2038 nir_block
*end_block
;
2040 /** list for all local variables in the function */
2041 struct exec_list locals
;
2043 /** list of local registers in the function */
2044 struct exec_list registers
;
2046 /** next available local register index */
2049 /** next available SSA value index */
2052 /* total number of basic blocks, only valid when block_index_dirty = false */
2053 unsigned num_blocks
;
2055 nir_metadata valid_metadata
;
2056 } nir_function_impl
;
2058 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2059 nir_start_block(nir_function_impl
*impl
)
2061 return (nir_block
*) impl
->body
.head_sentinel
.next
;
2064 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2065 nir_impl_last_block(nir_function_impl
*impl
)
2067 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
2070 static inline nir_cf_node
*
2071 nir_cf_node_next(nir_cf_node
*node
)
2073 struct exec_node
*next
= exec_node_get_next(&node
->node
);
2074 if (exec_node_is_tail_sentinel(next
))
2077 return exec_node_data(nir_cf_node
, next
, node
);
2080 static inline nir_cf_node
*
2081 nir_cf_node_prev(nir_cf_node
*node
)
2083 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
2084 if (exec_node_is_head_sentinel(prev
))
2087 return exec_node_data(nir_cf_node
, prev
, node
);
2091 nir_cf_node_is_first(const nir_cf_node
*node
)
2093 return exec_node_is_head_sentinel(node
->node
.prev
);
2097 nir_cf_node_is_last(const nir_cf_node
*node
)
2099 return exec_node_is_tail_sentinel(node
->node
.next
);
2102 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
2103 type
, nir_cf_node_block
)
2104 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
2105 type
, nir_cf_node_if
)
2106 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
2107 type
, nir_cf_node_loop
)
2108 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
2109 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
2111 static inline nir_block
*
2112 nir_if_first_then_block(nir_if
*if_stmt
)
2114 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
2115 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2118 static inline nir_block
*
2119 nir_if_last_then_block(nir_if
*if_stmt
)
2121 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
2122 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2125 static inline nir_block
*
2126 nir_if_first_else_block(nir_if
*if_stmt
)
2128 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
2129 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2132 static inline nir_block
*
2133 nir_if_last_else_block(nir_if
*if_stmt
)
2135 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
2136 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2139 static inline nir_block
*
2140 nir_loop_first_block(nir_loop
*loop
)
2142 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
2143 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2146 static inline nir_block
*
2147 nir_loop_last_block(nir_loop
*loop
)
2149 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
2150 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2154 * Return true if this list of cf_nodes contains a single empty block.
2157 nir_cf_list_is_empty_block(struct exec_list
*cf_list
)
2159 if (exec_list_is_singular(cf_list
)) {
2160 struct exec_node
*head
= exec_list_get_head(cf_list
);
2162 nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2163 return exec_list_is_empty(&block
->instr_list
);
2169 uint8_t num_components
;
2173 typedef struct nir_function
{
2174 struct exec_node node
;
2177 struct nir_shader
*shader
;
2179 unsigned num_params
;
2180 nir_parameter
*params
;
2182 /** The implementation of this function.
2184 * If the function is only declared and not implemented, this is NULL.
2186 nir_function_impl
*impl
;
2192 nir_lower_imul64
= (1 << 0),
2193 nir_lower_isign64
= (1 << 1),
2194 /** Lower all int64 modulus and division opcodes */
2195 nir_lower_divmod64
= (1 << 2),
2196 /** Lower all 64-bit umul_high and imul_high opcodes */
2197 nir_lower_imul_high64
= (1 << 3),
2198 nir_lower_mov64
= (1 << 4),
2199 nir_lower_icmp64
= (1 << 5),
2200 nir_lower_iadd64
= (1 << 6),
2201 nir_lower_iabs64
= (1 << 7),
2202 nir_lower_ineg64
= (1 << 8),
2203 nir_lower_logic64
= (1 << 9),
2204 nir_lower_minmax64
= (1 << 10),
2205 nir_lower_shift64
= (1 << 11),
2206 nir_lower_imul_2x32_64
= (1 << 12),
2207 } nir_lower_int64_options
;
2210 nir_lower_drcp
= (1 << 0),
2211 nir_lower_dsqrt
= (1 << 1),
2212 nir_lower_drsq
= (1 << 2),
2213 nir_lower_dtrunc
= (1 << 3),
2214 nir_lower_dfloor
= (1 << 4),
2215 nir_lower_dceil
= (1 << 5),
2216 nir_lower_dfract
= (1 << 6),
2217 nir_lower_dround_even
= (1 << 7),
2218 nir_lower_dmod
= (1 << 8),
2219 nir_lower_fp64_full_software
= (1 << 9),
2220 } nir_lower_doubles_options
;
2222 typedef struct nir_shader_compiler_options
{
2228 /** Lowers flrp when it does not support doubles */
2236 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2237 bool lower_bitfield_extract
;
2238 /** Lowers ibitfield_extract/ubitfield_extract to bfm, compares, shifts. */
2239 bool lower_bitfield_extract_to_shifts
;
2240 /** Lowers bitfield_insert to bfi/bfm */
2241 bool lower_bitfield_insert
;
2242 /** Lowers bitfield_insert to bfm, compares, and shifts. */
2243 bool lower_bitfield_insert_to_shifts
;
2244 /** Lowers bitfield_reverse to shifts. */
2245 bool lower_bitfield_reverse
;
2246 /** Lowers bit_count to shifts. */
2247 bool lower_bit_count
;
2248 /** Lowers bfm to shifts and subtracts. */
2250 /** Lowers ifind_msb to compare and ufind_msb */
2251 bool lower_ifind_msb
;
2252 /** Lowers find_lsb to ufind_msb and logic ops */
2253 bool lower_find_lsb
;
2254 bool lower_uadd_carry
;
2255 bool lower_usub_borrow
;
2256 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2257 bool lower_mul_high
;
2258 /** lowers fneg and ineg to fsub and isub. */
2260 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2263 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2266 /** enables rules to lower idiv by power-of-two: */
2269 /** enables rules to lower isign to imin+imax */
2272 /** enables rules to lower fsign to fsub and flt */
2275 /* Does the native fdot instruction replicate its result for four
2276 * components? If so, then opt_algebraic_late will turn all fdotN
2277 * instructions into fdot_replicatedN instructions.
2279 bool fdot_replicates
;
2281 /** lowers ffloor to fsub+ffract: */
2284 /** lowers ffract to fsub+ffloor: */
2287 /** lowers fceil to fneg+ffloor+fneg: */
2294 bool lower_pack_half_2x16
;
2295 bool lower_pack_unorm_2x16
;
2296 bool lower_pack_snorm_2x16
;
2297 bool lower_pack_unorm_4x8
;
2298 bool lower_pack_snorm_4x8
;
2299 bool lower_unpack_half_2x16
;
2300 bool lower_unpack_unorm_2x16
;
2301 bool lower_unpack_snorm_2x16
;
2302 bool lower_unpack_unorm_4x8
;
2303 bool lower_unpack_snorm_4x8
;
2305 bool lower_extract_byte
;
2306 bool lower_extract_word
;
2308 bool lower_all_io_to_temps
;
2309 bool lower_all_io_to_elements
;
2311 /* Indicates that the driver only has zero-based vertex id */
2312 bool vertex_id_zero_based
;
2315 * If enabled, gl_BaseVertex will be lowered as:
2316 * is_indexed_draw (~0/0) & firstvertex
2318 bool lower_base_vertex
;
2321 * If enabled, gl_HelperInvocation will be lowered as:
2323 * !((1 << sample_id) & sample_mask_in))
2325 * This depends on some possibly hw implementation details, which may
2326 * not be true for all hw. In particular that the FS is only executed
2327 * for covered samples or for helper invocations. So, do not blindly
2328 * enable this option.
2330 * Note: See also issue #22 in ARB_shader_image_load_store
2332 bool lower_helper_invocation
;
2335 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
2337 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
2338 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
2340 bool optimize_sample_mask_in
;
2342 bool lower_cs_local_index_from_id
;
2343 bool lower_cs_local_id_from_index
;
2345 bool lower_device_index_to_zero
;
2347 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
2348 bool lower_wpos_pntc
;
2354 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
2355 * for IO purposes and would prefer loads/stores be vectorized.
2360 * Should nir_lower_io() create load_interpolated_input intrinsics?
2362 * If not, it generates regular load_input intrinsics and interpolation
2363 * information must be inferred from the list of input nir_variables.
2365 bool use_interpolated_input_intrinsics
;
2367 /* Lowers when 32x32->64 bit multiplication is not supported */
2368 bool lower_mul_2x32_64
;
2370 unsigned max_unroll_iterations
;
2372 nir_lower_int64_options lower_int64_options
;
2373 nir_lower_doubles_options lower_doubles_options
;
2374 } nir_shader_compiler_options
;
2376 typedef struct nir_shader
{
2377 /** list of uniforms (nir_variable) */
2378 struct exec_list uniforms
;
2380 /** list of inputs (nir_variable) */
2381 struct exec_list inputs
;
2383 /** list of outputs (nir_variable) */
2384 struct exec_list outputs
;
2386 /** list of shared compute variables (nir_variable) */
2387 struct exec_list shared
;
2389 /** Set of driver-specific options for the shader.
2391 * The memory for the options is expected to be kept in a single static
2392 * copy by the driver.
2394 const struct nir_shader_compiler_options
*options
;
2396 /** Various bits of compile-time information about a given shader */
2397 struct shader_info info
;
2399 /** list of global variables in the shader (nir_variable) */
2400 struct exec_list globals
;
2402 /** list of system value variables in the shader (nir_variable) */
2403 struct exec_list system_values
;
2405 struct exec_list functions
; /** < list of nir_function */
2408 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
2411 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
2413 /** Size in bytes of required scratch space */
2414 unsigned scratch_size
;
2416 /** Constant data associated with this shader.
2418 * Constant data is loaded through load_constant intrinsics. See also
2419 * nir_opt_large_constants.
2421 void *constant_data
;
2422 unsigned constant_data_size
;
2425 #define nir_foreach_function(func, shader) \
2426 foreach_list_typed(nir_function, func, node, &(shader)->functions)
2428 static inline nir_function_impl
*
2429 nir_shader_get_entrypoint(nir_shader
*shader
)
2431 nir_function
*func
= NULL
;
2433 nir_foreach_function(function
, shader
) {
2434 assert(func
== NULL
);
2435 if (function
->is_entrypoint
) {
2446 assert(func
->num_params
== 0);
2451 nir_shader
*nir_shader_create(void *mem_ctx
,
2452 gl_shader_stage stage
,
2453 const nir_shader_compiler_options
*options
,
2456 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
2458 void nir_reg_remove(nir_register
*reg
);
2460 /** Adds a variable to the appropriate list in nir_shader */
2461 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
2464 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
2466 assert(var
->data
.mode
== nir_var_function_temp
);
2467 exec_list_push_tail(&impl
->locals
, &var
->node
);
2470 /** creates a variable, sets a few defaults, and adds it to the list */
2471 nir_variable
*nir_variable_create(nir_shader
*shader
,
2472 nir_variable_mode mode
,
2473 const struct glsl_type
*type
,
2475 /** creates a local variable and adds it to the list */
2476 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
2477 const struct glsl_type
*type
,
2480 /** creates a function and adds it to the shader's list of functions */
2481 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
2483 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
2484 /** creates a function_impl that isn't tied to any particular function */
2485 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
2487 nir_block
*nir_block_create(nir_shader
*shader
);
2488 nir_if
*nir_if_create(nir_shader
*shader
);
2489 nir_loop
*nir_loop_create(nir_shader
*shader
);
2491 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
2493 /** requests that the given pieces of metadata be generated */
2494 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
2495 /** dirties all but the preserved metadata */
2496 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
2498 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
2499 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
2501 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
2502 nir_deref_type deref_type
);
2504 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
2506 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
2507 unsigned num_components
,
2510 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
2511 nir_intrinsic_op op
);
2513 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
2514 nir_function
*callee
);
2516 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
2518 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
2520 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
2522 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
2523 unsigned num_components
,
2526 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
2529 * NIR Cursors and Instruction Insertion API
2532 * A tiny struct representing a point to insert/extract instructions or
2533 * control flow nodes. Helps reduce the combinatorial explosion of possible
2534 * points to insert/extract.
2536 * \sa nir_control_flow.h
2539 nir_cursor_before_block
,
2540 nir_cursor_after_block
,
2541 nir_cursor_before_instr
,
2542 nir_cursor_after_instr
,
2543 } nir_cursor_option
;
2546 nir_cursor_option option
;
2553 static inline nir_block
*
2554 nir_cursor_current_block(nir_cursor cursor
)
2556 if (cursor
.option
== nir_cursor_before_instr
||
2557 cursor
.option
== nir_cursor_after_instr
) {
2558 return cursor
.instr
->block
;
2560 return cursor
.block
;
2564 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
2566 static inline nir_cursor
2567 nir_before_block(nir_block
*block
)
2570 cursor
.option
= nir_cursor_before_block
;
2571 cursor
.block
= block
;
2575 static inline nir_cursor
2576 nir_after_block(nir_block
*block
)
2579 cursor
.option
= nir_cursor_after_block
;
2580 cursor
.block
= block
;
2584 static inline nir_cursor
2585 nir_before_instr(nir_instr
*instr
)
2588 cursor
.option
= nir_cursor_before_instr
;
2589 cursor
.instr
= instr
;
2593 static inline nir_cursor
2594 nir_after_instr(nir_instr
*instr
)
2597 cursor
.option
= nir_cursor_after_instr
;
2598 cursor
.instr
= instr
;
2602 static inline nir_cursor
2603 nir_after_block_before_jump(nir_block
*block
)
2605 nir_instr
*last_instr
= nir_block_last_instr(block
);
2606 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
2607 return nir_before_instr(last_instr
);
2609 return nir_after_block(block
);
2613 static inline nir_cursor
2614 nir_before_src(nir_src
*src
, bool is_if_condition
)
2616 if (is_if_condition
) {
2617 nir_block
*prev_block
=
2618 nir_cf_node_as_block(nir_cf_node_prev(&src
->parent_if
->cf_node
));
2619 assert(!nir_block_ends_in_jump(prev_block
));
2620 return nir_after_block(prev_block
);
2621 } else if (src
->parent_instr
->type
== nir_instr_type_phi
) {
2623 nir_phi_instr
*cond_phi
= nir_instr_as_phi(src
->parent_instr
);
2625 nir_foreach_phi_src(phi_src
, cond_phi
) {
2626 if (phi_src
->src
.ssa
== src
->ssa
) {
2633 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
2634 * to have a more specific name.
2636 nir_phi_src
*phi_src
= LIST_ENTRY(nir_phi_src
, src
, src
);
2637 return nir_after_block_before_jump(phi_src
->pred
);
2639 return nir_before_instr(src
->parent_instr
);
2643 static inline nir_cursor
2644 nir_before_cf_node(nir_cf_node
*node
)
2646 if (node
->type
== nir_cf_node_block
)
2647 return nir_before_block(nir_cf_node_as_block(node
));
2649 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
2652 static inline nir_cursor
2653 nir_after_cf_node(nir_cf_node
*node
)
2655 if (node
->type
== nir_cf_node_block
)
2656 return nir_after_block(nir_cf_node_as_block(node
));
2658 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
2661 static inline nir_cursor
2662 nir_after_phis(nir_block
*block
)
2664 nir_foreach_instr(instr
, block
) {
2665 if (instr
->type
!= nir_instr_type_phi
)
2666 return nir_before_instr(instr
);
2668 return nir_after_block(block
);
2671 static inline nir_cursor
2672 nir_after_cf_node_and_phis(nir_cf_node
*node
)
2674 if (node
->type
== nir_cf_node_block
)
2675 return nir_after_block(nir_cf_node_as_block(node
));
2677 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
2679 return nir_after_phis(block
);
2682 static inline nir_cursor
2683 nir_before_cf_list(struct exec_list
*cf_list
)
2685 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
2686 exec_list_get_head(cf_list
), node
);
2687 return nir_before_cf_node(first_node
);
2690 static inline nir_cursor
2691 nir_after_cf_list(struct exec_list
*cf_list
)
2693 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
2694 exec_list_get_tail(cf_list
), node
);
2695 return nir_after_cf_node(last_node
);
2699 * Insert a NIR instruction at the given cursor.
2701 * Note: This does not update the cursor.
2703 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
2706 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
2708 nir_instr_insert(nir_before_instr(instr
), before
);
2712 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
2714 nir_instr_insert(nir_after_instr(instr
), after
);
2718 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
2720 nir_instr_insert(nir_before_block(block
), before
);
2724 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
2726 nir_instr_insert(nir_after_block(block
), after
);
2730 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
2732 nir_instr_insert(nir_before_cf_node(node
), before
);
2736 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
2738 nir_instr_insert(nir_after_cf_node(node
), after
);
2742 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
2744 nir_instr_insert(nir_before_cf_list(list
), before
);
2748 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
2750 nir_instr_insert(nir_after_cf_list(list
), after
);
2753 void nir_instr_remove_v(nir_instr
*instr
);
2755 static inline nir_cursor
2756 nir_instr_remove(nir_instr
*instr
)
2759 nir_instr
*prev
= nir_instr_prev(instr
);
2761 cursor
= nir_after_instr(prev
);
2763 cursor
= nir_before_block(instr
->block
);
2765 nir_instr_remove_v(instr
);
2771 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
2772 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
2773 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
2774 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
2776 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
2777 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
2779 nir_const_value
*nir_src_as_const_value(nir_src src
);
2781 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
2782 static inline c_type * \
2783 nir_src_as_ ## name (nir_src src) \
2785 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
2786 ? cast_macro(src.ssa->parent_instr) : NULL; \
2789 NIR_SRC_AS_(alu_instr
, nir_alu_instr
, nir_instr_type_alu
, nir_instr_as_alu
)
2790 NIR_SRC_AS_(intrinsic
, nir_intrinsic_instr
,
2791 nir_instr_type_intrinsic
, nir_instr_as_intrinsic
)
2792 NIR_SRC_AS_(deref
, nir_deref_instr
, nir_instr_type_deref
, nir_instr_as_deref
)
2794 bool nir_src_is_dynamically_uniform(nir_src src
);
2795 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
2796 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
2797 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
2798 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
2799 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
2802 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
2803 unsigned num_components
, unsigned bit_size
,
2805 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
2806 unsigned num_components
, unsigned bit_size
,
2809 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
2810 const struct glsl_type
*type
,
2813 assert(glsl_type_is_vector_or_scalar(type
));
2814 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
2815 glsl_get_bit_size(type
), name
);
2817 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
2818 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
2819 nir_instr
*after_me
);
2821 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
2824 * finds the next basic block in source-code order, returns NULL if there is
2828 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
2830 /* Performs the opposite of nir_block_cf_tree_next() */
2832 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
2834 /* Gets the first block in a CF node in source-code order */
2836 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
2838 /* Gets the last block in a CF node in source-code order */
2840 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
2842 /* Gets the next block after a CF node in source-code order */
2844 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
2846 /* Macros for loops that visit blocks in source-code order */
2848 #define nir_foreach_block(block, impl) \
2849 for (nir_block *block = nir_start_block(impl); block != NULL; \
2850 block = nir_block_cf_tree_next(block))
2852 #define nir_foreach_block_safe(block, impl) \
2853 for (nir_block *block = nir_start_block(impl), \
2854 *next = nir_block_cf_tree_next(block); \
2856 block = next, next = nir_block_cf_tree_next(block))
2858 #define nir_foreach_block_reverse(block, impl) \
2859 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
2860 block = nir_block_cf_tree_prev(block))
2862 #define nir_foreach_block_reverse_safe(block, impl) \
2863 for (nir_block *block = nir_impl_last_block(impl), \
2864 *prev = nir_block_cf_tree_prev(block); \
2866 block = prev, prev = nir_block_cf_tree_prev(block))
2868 #define nir_foreach_block_in_cf_node(block, node) \
2869 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
2870 block != nir_cf_node_cf_tree_next(node); \
2871 block = nir_block_cf_tree_next(block))
2873 /* If the following CF node is an if, this function returns that if.
2874 * Otherwise, it returns NULL.
2876 nir_if
*nir_block_get_following_if(nir_block
*block
);
2878 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
2880 void nir_index_local_regs(nir_function_impl
*impl
);
2881 void nir_index_ssa_defs(nir_function_impl
*impl
);
2882 unsigned nir_index_instrs(nir_function_impl
*impl
);
2884 void nir_index_blocks(nir_function_impl
*impl
);
2886 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
2887 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
2888 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
2889 void nir_print_deref(const nir_deref_instr
*deref
, FILE *fp
);
2891 /** Shallow clone of a single ALU instruction. */
2892 nir_alu_instr
*nir_alu_instr_clone(nir_shader
*s
, const nir_alu_instr
*orig
);
2894 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
2895 nir_function_impl
*nir_function_impl_clone(nir_shader
*shader
,
2896 const nir_function_impl
*fi
);
2897 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
2898 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
2900 nir_shader
*nir_shader_serialize_deserialize(void *mem_ctx
, nir_shader
*s
);
2903 void nir_validate_shader(nir_shader
*shader
, const char *when
);
2904 void nir_metadata_set_validation_flag(nir_shader
*shader
);
2905 void nir_metadata_check_validation_flag(nir_shader
*shader
);
2908 should_skip_nir(const char *name
)
2910 static const char *list
= NULL
;
2912 /* Comma separated list of names to skip. */
2913 list
= getenv("NIR_SKIP");
2921 return comma_separated_list_contains(list
, name
);
2925 should_clone_nir(void)
2927 static int should_clone
= -1;
2928 if (should_clone
< 0)
2929 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
2931 return should_clone
;
2935 should_serialize_deserialize_nir(void)
2937 static int test_serialize
= -1;
2938 if (test_serialize
< 0)
2939 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
2941 return test_serialize
;
2945 should_print_nir(void)
2947 static int should_print
= -1;
2948 if (should_print
< 0)
2949 should_print
= env_var_as_boolean("NIR_PRINT", false);
2951 return should_print
;
2954 static inline void nir_validate_shader(nir_shader
*shader
, const char *when
) { (void) shader
; (void)when
; }
2955 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
2956 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
2957 static inline bool should_skip_nir(UNUSED
const char *pass_name
) { return false; }
2958 static inline bool should_clone_nir(void) { return false; }
2959 static inline bool should_serialize_deserialize_nir(void) { return false; }
2960 static inline bool should_print_nir(void) { return false; }
2963 #define _PASS(pass, nir, do_pass) do { \
2964 if (should_skip_nir(#pass)) { \
2965 printf("skipping %s\n", #pass); \
2969 nir_validate_shader(nir, "after " #pass); \
2970 if (should_clone_nir()) { \
2971 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
2975 if (should_serialize_deserialize_nir()) { \
2976 void *mem_ctx = ralloc_parent(nir); \
2977 nir = nir_shader_serialize_deserialize(mem_ctx, nir); \
2981 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
2982 nir_metadata_set_validation_flag(nir); \
2983 if (should_print_nir()) \
2984 printf("%s\n", #pass); \
2985 if (pass(nir, ##__VA_ARGS__)) { \
2987 if (should_print_nir()) \
2988 nir_print_shader(nir, stdout); \
2989 nir_metadata_check_validation_flag(nir); \
2993 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
2994 if (should_print_nir()) \
2995 printf("%s\n", #pass); \
2996 pass(nir, ##__VA_ARGS__); \
2997 if (should_print_nir()) \
2998 nir_print_shader(nir, stdout); \
3001 #define NIR_SKIP(name) should_skip_nir(#name)
3003 void nir_calc_dominance_impl(nir_function_impl
*impl
);
3004 void nir_calc_dominance(nir_shader
*shader
);
3006 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
3007 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
3009 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
3010 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
3012 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
3013 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
3015 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
3016 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
3018 int nir_gs_count_vertices(const nir_shader
*shader
);
3020 bool nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3021 bool nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3022 bool nir_split_var_copies(nir_shader
*shader
);
3023 bool nir_split_per_member_structs(nir_shader
*shader
);
3024 bool nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
);
3026 bool nir_lower_returns_impl(nir_function_impl
*impl
);
3027 bool nir_lower_returns(nir_shader
*shader
);
3029 void nir_inline_function_impl(struct nir_builder
*b
,
3030 const nir_function_impl
*impl
,
3031 nir_ssa_def
**params
);
3032 bool nir_inline_functions(nir_shader
*shader
);
3034 bool nir_propagate_invariant(nir_shader
*shader
);
3036 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
3037 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
3038 nir_intrinsic_instr
*copy
);
3039 bool nir_lower_var_copies(nir_shader
*shader
);
3041 void nir_fixup_deref_modes(nir_shader
*shader
);
3043 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
3046 nir_lower_direct_array_deref_of_vec_load
= (1 << 0),
3047 nir_lower_indirect_array_deref_of_vec_load
= (1 << 1),
3048 nir_lower_direct_array_deref_of_vec_store
= (1 << 2),
3049 nir_lower_indirect_array_deref_of_vec_store
= (1 << 3),
3050 } nir_lower_array_deref_of_vec_options
;
3052 bool nir_lower_array_deref_of_vec(nir_shader
*shader
, nir_variable_mode modes
,
3053 nir_lower_array_deref_of_vec_options options
);
3055 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
3057 bool nir_lower_locals_to_regs(nir_shader
*shader
);
3059 void nir_lower_io_to_temporaries(nir_shader
*shader
,
3060 nir_function_impl
*entrypoint
,
3061 bool outputs
, bool inputs
);
3063 bool nir_lower_vars_to_scratch(nir_shader
*shader
,
3064 nir_variable_mode modes
,
3066 glsl_type_size_align_func size_align
);
3068 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
3070 void nir_gather_ssa_types(nir_function_impl
*impl
,
3071 BITSET_WORD
*float_types
,
3072 BITSET_WORD
*int_types
);
3074 void nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
3075 int (*type_size
)(const struct glsl_type
*, bool));
3077 /* Some helpers to do very simple linking */
3078 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3079 bool nir_remove_unused_io_vars(nir_shader
*shader
, struct exec_list
*var_list
,
3080 uint64_t *used_by_other_stage
,
3081 uint64_t *used_by_other_stage_patches
);
3082 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
3083 bool default_to_smooth_interp
);
3084 void nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3085 bool nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3088 /* If set, this forces all non-flat fragment shader inputs to be
3089 * interpolated as if with the "sample" qualifier. This requires
3090 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
3092 nir_lower_io_force_sample_interpolation
= (1 << 1),
3093 } nir_lower_io_options
;
3094 bool nir_lower_io(nir_shader
*shader
,
3095 nir_variable_mode modes
,
3096 int (*type_size
)(const struct glsl_type
*, bool),
3097 nir_lower_io_options
);
3101 * An address format which is a simple 32-bit global GPU address.
3103 nir_address_format_32bit_global
,
3106 * An address format which is a simple 64-bit global GPU address.
3108 nir_address_format_64bit_global
,
3111 * An address format which is a bounds-checked 64-bit global GPU address.
3113 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
3114 * address stored with the low bits in .x and high bits in .y, .z is a
3115 * size, and .w is an offset. When the final I/O operation is lowered, .w
3116 * is checked against .z and the operation is predicated on the result.
3118 nir_address_format_64bit_bounded_global
,
3121 * An address format which is comprised of a vec2 where the first
3122 * component is a buffer index and the second is an offset.
3124 nir_address_format_32bit_index_offset
,
3127 * An address format which is a simple 32-bit offset.
3129 nir_address_format_32bit_offset
,
3132 * An address format representing a purely logical addressing model. In
3133 * this model, all deref chains must be complete from the dereference
3134 * operation to the variable. Cast derefs are not allowed. These
3135 * addresses will be 32-bit scalars but the format is immaterial because
3136 * you can always chase the chain.
3138 nir_address_format_logical
,
3139 } nir_address_format
;
3141 static inline unsigned
3142 nir_address_format_bit_size(nir_address_format addr_format
)
3144 switch (addr_format
) {
3145 case nir_address_format_32bit_global
: return 32;
3146 case nir_address_format_64bit_global
: return 64;
3147 case nir_address_format_64bit_bounded_global
: return 32;
3148 case nir_address_format_32bit_index_offset
: return 32;
3149 case nir_address_format_32bit_offset
: return 32;
3150 case nir_address_format_logical
: return 32;
3152 unreachable("Invalid address format");
3155 static inline unsigned
3156 nir_address_format_num_components(nir_address_format addr_format
)
3158 switch (addr_format
) {
3159 case nir_address_format_32bit_global
: return 1;
3160 case nir_address_format_64bit_global
: return 1;
3161 case nir_address_format_64bit_bounded_global
: return 4;
3162 case nir_address_format_32bit_index_offset
: return 2;
3163 case nir_address_format_32bit_offset
: return 1;
3164 case nir_address_format_logical
: return 1;
3166 unreachable("Invalid address format");
3169 static inline const struct glsl_type
*
3170 nir_address_format_to_glsl_type(nir_address_format addr_format
)
3172 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
3173 assert(bit_size
== 32 || bit_size
== 64);
3174 return glsl_vector_type(bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
,
3175 nir_address_format_num_components(addr_format
));
3178 const nir_const_value
*nir_address_format_null_value(nir_address_format addr_format
);
3180 nir_ssa_def
* nir_explicit_io_address_from_deref(struct nir_builder
*b
,
3181 nir_deref_instr
*deref
,
3182 nir_ssa_def
*base_addr
,
3183 nir_address_format addr_format
);
3184 void nir_lower_explicit_io_instr(struct nir_builder
*b
,
3185 nir_intrinsic_instr
*io_instr
,
3187 nir_address_format addr_format
);
3189 bool nir_lower_explicit_io(nir_shader
*shader
,
3190 nir_variable_mode modes
,
3191 nir_address_format
);
3193 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
3194 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
3196 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
3198 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
3199 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
3200 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
3202 bool nir_remove_dead_derefs(nir_shader
*shader
);
3203 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
3204 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
);
3205 bool nir_lower_constant_initializers(nir_shader
*shader
,
3206 nir_variable_mode modes
);
3208 bool nir_move_load_const(nir_shader
*shader
);
3209 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
3210 bool nir_lower_vec_to_movs(nir_shader
*shader
);
3211 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
3213 bool nir_lower_alu(nir_shader
*shader
);
3215 bool nir_lower_flrp(nir_shader
*shader
, unsigned lowering_mask
,
3216 bool always_precise
, bool have_ffma
);
3218 bool nir_lower_alu_to_scalar(nir_shader
*shader
, BITSET_WORD
*lower_set
);
3219 bool nir_lower_bool_to_float(nir_shader
*shader
);
3220 bool nir_lower_bool_to_int32(nir_shader
*shader
);
3221 bool nir_lower_int_to_float(nir_shader
*shader
);
3222 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
3223 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
3224 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
3225 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
3226 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
3228 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
3229 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
3230 bool nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode mask
);
3232 void nir_lower_fragcoord_wtrans(nir_shader
*shader
);
3233 void nir_lower_viewport_transform(nir_shader
*shader
);
3234 bool nir_lower_uniforms_to_ubo(nir_shader
*shader
, int multiplier
);
3236 typedef struct nir_lower_subgroups_options
{
3237 uint8_t subgroup_size
;
3238 uint8_t ballot_bit_size
;
3239 bool lower_to_scalar
:1;
3240 bool lower_vote_trivial
:1;
3241 bool lower_vote_eq_to_ballot
:1;
3242 bool lower_subgroup_masks
:1;
3243 bool lower_shuffle
:1;
3244 bool lower_shuffle_to_32bit
:1;
3246 } nir_lower_subgroups_options
;
3248 bool nir_lower_subgroups(nir_shader
*shader
,
3249 const nir_lower_subgroups_options
*options
);
3251 bool nir_lower_system_values(nir_shader
*shader
);
3253 enum PACKED nir_lower_tex_packing
{
3254 nir_lower_tex_packing_none
= 0,
3255 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
3256 * or unsigned ints based on the sampler type
3258 nir_lower_tex_packing_16
,
3259 /* The sampler returns 1 32-bit word of 4x8 unorm */
3260 nir_lower_tex_packing_8
,
3263 typedef struct nir_lower_tex_options
{
3265 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
3266 * sampler types a texture projector is lowered.
3271 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
3273 bool lower_txf_offset
;
3276 * If true, lower away nir_tex_src_offset for all rect textures.
3278 bool lower_rect_offset
;
3281 * If true, lower rect textures to 2D, using txs to fetch the
3282 * texture dimensions and dividing the texture coords by the
3283 * texture dims to normalize.
3288 * If true, convert yuv to rgb.
3290 unsigned lower_y_uv_external
;
3291 unsigned lower_y_u_v_external
;
3292 unsigned lower_yx_xuxv_external
;
3293 unsigned lower_xy_uxvx_external
;
3294 unsigned lower_ayuv_external
;
3295 unsigned lower_xyuv_external
;
3298 * To emulate certain texture wrap modes, this can be used
3299 * to saturate the specified tex coord to [0.0, 1.0]. The
3300 * bits are according to sampler #, ie. if, for example:
3302 * (conf->saturate_s & (1 << n))
3304 * is true, then the s coord for sampler n is saturated.
3306 * Note that clamping must happen *after* projector lowering
3307 * so any projected texture sample instruction with a clamped
3308 * coordinate gets automatically lowered, regardless of the
3309 * 'lower_txp' setting.
3311 unsigned saturate_s
;
3312 unsigned saturate_t
;
3313 unsigned saturate_r
;
3315 /* Bitmask of textures that need swizzling.
3317 * If (swizzle_result & (1 << texture_index)), then the swizzle in
3318 * swizzles[texture_index] is applied to the result of the texturing
3321 unsigned swizzle_result
;
3323 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
3324 * while 4 and 5 represent 0 and 1 respectively.
3326 uint8_t swizzles
[32][4];
3328 /* Can be used to scale sampled values in range required by the format. */
3329 float scale_factors
[32];
3332 * Bitmap of textures that need srgb to linear conversion. If
3333 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
3334 * of the texture are lowered to linear.
3336 unsigned lower_srgb
;
3339 * If true, lower nir_texop_tex on shaders that doesn't support implicit
3340 * LODs to nir_texop_txl.
3342 bool lower_tex_without_implicit_lod
;
3345 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
3347 bool lower_txd_cube_map
;
3350 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
3355 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
3356 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
3357 * with lower_txd_cube_map.
3359 bool lower_txd_shadow
;
3362 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
3363 * Implies lower_txd_cube_map and lower_txd_shadow.
3368 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
3369 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
3371 bool lower_txb_shadow_clamp
;
3374 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
3375 * with nir_texop_txl. This includes cube maps.
3377 bool lower_txd_shadow_clamp
;
3380 * If true, lower nir_texop_txd on when it uses both offset and min_lod
3381 * with nir_texop_txl. This includes cube maps.
3383 bool lower_txd_offset_clamp
;
3386 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
3387 * sampler is bindless.
3389 bool lower_txd_clamp_bindless_sampler
;
3392 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
3393 * sampler index is not statically determinable to be less than 16.
3395 bool lower_txd_clamp_if_sampler_index_not_lt_16
;
3398 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
3399 * mixed-up tg4 locations.
3401 bool lower_tg4_broadcom_swizzle
;
3404 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
3406 bool lower_tg4_offsets
;
3408 enum nir_lower_tex_packing lower_tex_packing
[32];
3409 } nir_lower_tex_options
;
3411 bool nir_lower_tex(nir_shader
*shader
,
3412 const nir_lower_tex_options
*options
);
3414 enum nir_lower_non_uniform_access_type
{
3415 nir_lower_non_uniform_ubo_access
= (1 << 0),
3416 nir_lower_non_uniform_ssbo_access
= (1 << 1),
3417 nir_lower_non_uniform_texture_access
= (1 << 2),
3418 nir_lower_non_uniform_image_access
= (1 << 3),
3421 bool nir_lower_non_uniform_access(nir_shader
*shader
,
3422 enum nir_lower_non_uniform_access_type
);
3424 bool nir_lower_idiv(nir_shader
*shader
);
3426 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
, bool use_vars
);
3427 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
);
3428 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
3430 bool nir_lower_frexp(nir_shader
*nir
);
3432 void nir_lower_two_sided_color(nir_shader
*shader
);
3434 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
3436 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
3437 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
3438 const gl_state_index16
*uniform_state_tokens
);
3440 typedef struct nir_lower_wpos_ytransform_options
{
3441 gl_state_index16 state_tokens
[STATE_LENGTH
];
3442 bool fs_coord_origin_upper_left
:1;
3443 bool fs_coord_origin_lower_left
:1;
3444 bool fs_coord_pixel_center_integer
:1;
3445 bool fs_coord_pixel_center_half_integer
:1;
3446 } nir_lower_wpos_ytransform_options
;
3448 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
3449 const nir_lower_wpos_ytransform_options
*options
);
3450 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
3452 bool nir_lower_fb_read(nir_shader
*shader
);
3454 typedef struct nir_lower_drawpixels_options
{
3455 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
3456 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
3457 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
3458 unsigned drawpix_sampler
;
3459 unsigned pixelmap_sampler
;
3461 bool scale_and_bias
:1;
3462 } nir_lower_drawpixels_options
;
3464 void nir_lower_drawpixels(nir_shader
*shader
,
3465 const nir_lower_drawpixels_options
*options
);
3467 typedef struct nir_lower_bitmap_options
{
3470 } nir_lower_bitmap_options
;
3472 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
3474 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
, unsigned ssbo_offset
);
3477 nir_lower_int_source_mods
= 1 << 0,
3478 nir_lower_float_source_mods
= 1 << 1,
3479 nir_lower_triop_abs
= 1 << 2,
3480 nir_lower_all_source_mods
= (1 << 3) - 1
3481 } nir_lower_to_source_mods_flags
;
3484 bool nir_lower_to_source_mods(nir_shader
*shader
, nir_lower_to_source_mods_flags options
);
3486 bool nir_lower_gs_intrinsics(nir_shader
*shader
);
3488 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
3490 bool nir_lower_bit_size(nir_shader
*shader
,
3491 nir_lower_bit_size_callback callback
,
3492 void *callback_data
);
3494 nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode
);
3495 bool nir_lower_int64(nir_shader
*shader
, nir_lower_int64_options options
);
3497 nir_lower_doubles_options
nir_lower_doubles_op_to_options_mask(nir_op opcode
);
3498 bool nir_lower_doubles(nir_shader
*shader
, const nir_shader
*softfp64
,
3499 nir_lower_doubles_options options
);
3500 bool nir_lower_pack(nir_shader
*shader
);
3502 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
3504 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
3506 void nir_loop_analyze_impl(nir_function_impl
*impl
,
3507 nir_variable_mode indirect_mask
);
3509 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
3511 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
3512 bool nir_repair_ssa(nir_shader
*shader
);
3514 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
3516 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
3517 * registers. If false, convert all values (even those not involved in a phi
3518 * node) to registers.
3520 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
3522 bool nir_lower_phis_to_regs_block(nir_block
*block
);
3523 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
3524 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
);
3526 bool nir_opt_comparison_pre(nir_shader
*shader
);
3528 bool nir_opt_algebraic(nir_shader
*shader
);
3529 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
3530 bool nir_opt_algebraic_late(nir_shader
*shader
);
3531 bool nir_opt_constant_folding(nir_shader
*shader
);
3533 bool nir_opt_combine_stores(nir_shader
*shader
, nir_variable_mode modes
);
3535 bool nir_copy_prop(nir_shader
*shader
);
3537 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
3539 bool nir_opt_cse(nir_shader
*shader
);
3541 bool nir_opt_dce(nir_shader
*shader
);
3543 bool nir_opt_dead_cf(nir_shader
*shader
);
3545 bool nir_opt_dead_write_vars(nir_shader
*shader
);
3547 bool nir_opt_deref_impl(nir_function_impl
*impl
);
3548 bool nir_opt_deref(nir_shader
*shader
);
3550 bool nir_opt_find_array_copies(nir_shader
*shader
);
3552 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
3554 bool nir_opt_idiv_const(nir_shader
*shader
, unsigned min_bit_size
);
3556 bool nir_opt_if(nir_shader
*shader
, bool aggressive_last_continue
);
3558 bool nir_opt_intrinsics(nir_shader
*shader
);
3560 bool nir_opt_large_constants(nir_shader
*shader
,
3561 glsl_type_size_align_func size_align
,
3562 unsigned threshold
);
3564 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
3566 bool nir_opt_move_comparisons(nir_shader
*shader
);
3568 bool nir_opt_move_load_ubo(nir_shader
*shader
);
3570 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
,
3571 bool indirect_load_ok
, bool expensive_alu_ok
);
3573 bool nir_opt_rematerialize_compares(nir_shader
*shader
);
3575 bool nir_opt_remove_phis(nir_shader
*shader
);
3577 bool nir_opt_shrink_load(nir_shader
*shader
);
3579 bool nir_opt_trivial_continues(nir_shader
*shader
);
3581 bool nir_opt_undef(nir_shader
*shader
);
3583 bool nir_opt_conditional_discard(nir_shader
*shader
);
3585 void nir_strip(nir_shader
*shader
);
3587 void nir_sweep(nir_shader
*shader
);
3589 void nir_remap_dual_slot_attributes(nir_shader
*shader
,
3590 uint64_t *dual_slot_inputs
);
3591 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
);
3593 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
3594 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);
3596 bool nir_lower_sincos(nir_shader
*shader
);